Skip to content

A lightweight WebAssembly runtime that is fast, secure, and standards-compliant

[Cranelift] resolve #12790, fix arithmetic rewrite for wider integers (#12796)

Hyunbin Kimbb5a845433f3
diff --git a/cranelift/codegen/src/opts/arithmetic.isle b/cranelift/codegen/src/opts/arithmetic.isle
index c6eba1aed1..fd5172f3d8 100644
--- a/cranelift/codegen/src/opts/arithmetic.isle
+++ b/cranelift/codegen/src/opts/arithmetic.isle
@@ -318,5 +318,9 @@
(rule (simplify (isub ty (bor ty y x) (bxor ty y x))) (band ty x y))
;; (~x) + x == -1
-(rule (simplify (iadd ty (bnot ty x) x)) (iconst_s ty -1))
-(rule (simplify (iadd ty x (bnot ty x))) (iconst_s ty -1))
+;; Keep the generic fold for <=64-bit types, and handle i128 explicitly.
+(rule (simplify (iadd (fits_in_64 ty) (bnot ty x) x)) (iconst_s ty -1))
+(rule (simplify (iadd (fits_in_64 ty) x (bnot ty x))) (iconst_s ty -1))
+
+(rule (simplify (iadd $I128 (bnot $I128 x) x)) (sextend $I128 (iconst_s $I64 -1)))
+(rule (simplify (iadd $I128 x (bnot $I128 x))) (sextend $I128 (iconst_s $I64 -1)))
diff --git a/cranelift/filetests/filetests/egraph/arithmetic-precise.clif b/cranelift/filetests/filetests/egraph/arithmetic-precise.clif
index 986150e594..f48b568548 100644
--- a/cranelift/filetests/filetests/egraph/arithmetic-precise.clif
+++ b/cranelift/filetests/filetests/egraph/arithmetic-precise.clif
@@ -71,3 +71,16 @@ block0(v0: i32):
; return v3 ; v3 = -1
; }
+function %fold_iadd_bnot_x_to_allones_i128(i128) -> i128 {
+block0(v0: i128):
+ v1 = bnot v0
+ v2 = iadd v1, v0
+ return v2
+}
+
+; function %fold_iadd_bnot_x_to_allones_i128(i128) -> i128 fast {
+; block0(v0: i128):
+; v3 = iconst.i64 -1
+; v4 = sextend.i128 v3 ; v3 = -1
+; return v4
+; }
\ No newline at end of file
diff --git a/cranelift/filetests/filetests/runtests/i128-arithmetic-extends.clif b/cranelift/filetests/filetests/runtests/i128-arithmetic-extends.clif
index 8e339797a7..4c42bc9bb5 100644
--- a/cranelift/filetests/filetests/runtests/i128-arithmetic-extends.clif
+++ b/cranelift/filetests/filetests/runtests/i128-arithmetic-extends.clif
@@ -1,6 +1,7 @@
test interpret
test run
set enable_llvm_abi_extensions=true
+set enable_multi_ret_implicit_sret
target aarch64
target s390x
target x86_64
@@ -17,3 +18,13 @@ block0(v0: i32, v1: i128):
; run: %sext_sshr_i32_i128(0x8000_0000, 0) == 0xFFFFFFFF80000000
; run: %sext_sshr_i32_i128(0x8000_0000, 32) == 0xFFFFFFFF80000000
; run: %sext_sshr_i32_i128(0x8000_0000, 0xFFFFFFFF_FFFFFFFF_FFFFFFFF_FFFFFF20) == 0xFFFFFFFF80000000
+
+function %fold_iadd_bnot_x_to_allones_i128_rt(i128) -> i128 {
+block0(v0: i128):
+ v1 = bnot v0
+ v2 = iadd v1, v0
+ return v2
+}
+; run: %fold_iadd_bnot_x_to_allones_i128_rt(0) == -1
+; run: %fold_iadd_bnot_x_to_allones_i128_rt(1) == -1
+; run: %fold_iadd_bnot_x_to_allones_i128_rt(0x12345678_9ABCDEF0_11111111_22222222) == -1